For the problem when using swiotlb

[Catalin Marinas]

On Wed, Nov 19, 2014 at 08:45:43AM +0000, Arnd Bergmann wrote:
> On Wednesday 19 November 2014 11:17:15 Ding Tianhong wrote:
> > On 2014/11/18 2:09, Catalin Marinas wrote:
> > > On Mon, Nov 17, 2014 at 12:18:42PM +0000, Arnd Bergmann wrote:
> > >> On Monday 17 November 2014 19:56:27 Ding Tianhong wrote:
> > >>>         The commit 3690951fc6d42f3a0903987677d0e592c49dd8db(arm64: Use swiotlb late initialisation)
> > >>> switches the DMA mapping code to swiotlb_tlb_late_init_with_default_size(), this will occur a problem
> > >>> when I run the scsi stress tests, the message as below:
> > >>>
> > >>>         sas_controller b1000000.sas: swiotlb buffer is full (sz: 65536 bytes)..
> > >>>         DMA: Out of SW-IOMMU space for 65536 bytes at device b1000000.sas
> > >>>
> > >>> The reason is that the swiotlb_tlb_late_init_with_default_size() could only alloc 16M memory for DMA-mapping,
> > >>> and the param in cmdline "swiotlb=xxx" is useless because the get_free_pages() only use the buddy to assigned a
> > >>> maximum memory of 16M(The MAX_ORDER is 13 for 4k pages), obviously 16M is too small in many scenes, but
> > >>> the swiotlb_init() which could reserved a bigger memory as wished could work well for most drivers.
> > >>>
> > >>> I could not get a better way to fix this problem except to revert this patch, so could you please give me some
> > >>> advise and help me, thanks very much.
> > >>
> > >> In general, you should not need to use swiotlb for most devices, in
> > >> particular for high-performance devices like network or block.
> > >>
> > >> Please make sure that you have set up the dma-ranges properties in
> > >> your DT properly to allow 64-bit DMA if the device supports it.
> > > 
> > > That's the problem indeed, the DMA API ends up using swiotlb bounce
> > > buffers because the physical address of the pages passed to (or
> > > allocated by) the driver are beyond 32-bit limit (which is the default
> > > dma mask).
> > > 
> > 
> > Thanks everyone, I think I found the way to fix it, need to enable DMA_CMA, to reserve a big memory
> > for CMA and set coherent mask for dev, then dma_alloc and dma_mapping will not use the swiotlb until
> > the memory out of mask or swiotlb_force is enabled.
> > 
> > If I still understand uncorrectly, please inform me.
> 
> Please do not use CMA to work around the problem, but fix the underlying bug
> instead.

Agree.

> The driver should call 'dma_set_mask_and_coherent()' with the appropriate
> dma mask, and check whether that succeeded. However, the code implementing
> dma_set_mask_and_coherent on arm64 also needs to be changed to look up
> the dma-ranges property (see of_dma_configure()), and check if the mask
> is possible.

dma_set_mask_and_coherent() is a generic function. I think the
of_dma_configure() should start with a coherent_dma_mask based on
dma-ranges if given rather than defaulting to DMA_BIT_MASK(32). The
comment in of_dma_configure() says that devices should set up the
supported mask but it's not always up to them but the bus they are
connected to.

Something like below, untested:

diff --git a/drivers/of/platform.c b/drivers/of/platform.c
index 3b64d0bf5bba..dff34883db45 100644
--- a/drivers/of/platform.c
+++ b/drivers/of/platform.c
@@ -200,6 +200,10 @@ static void of_dma_configure(struct device *dev)
 	/* DMA ranges found. Calculate and set dma_pfn_offset */
 	dev->dma_pfn_offset = PFN_DOWN(paddr - dma_addr);
 	dev_dbg(dev, "dma_pfn_offset(%#08lx)\n", dev->dma_pfn_offset);
+
+	/* Set the coherent_dma_mask based on the dma-ranges property */
+	if (size)
+		dev->coherent_dma_mask = DMA_BIT_MASK(ilog2(size));
 }
 
 /**